session-distill

name	session-distill
description	Distill Codex session JSONL files into review packets, local session notes, and reusable memory candidates. Use when the user asks to整理/提炼 Codex 对话, process `.codex/archived_sessions` or `.codex/sessions` rollouts, generate session packets, review packet coverage, or mark Codex sessions as distilled.

Codex Session Distiller

This is the Codex-native session-distill skill. It reads Codex rollout JSONL, not Claude Code session files.

Inputs And Outputs

• Input sessions:
  • ~/.codex/archived_sessions/rollout-*.jsonl
  • ~/.codex/sessions/**/*.jsonl
  • ~/.codex/session_index.jsonl for thread names when available
• Workspace:
  • ~/.codex/session-distill/manifest.json
  • ~/.codex/session-distill/packets/<session-id>.md
  • ~/.codex/session-distill/distilled/sessions/<session-id>.md
  • ~/.codex/session-distill/knowledge-base.md
  • optional ~/.codex/session-distill/memory-drafts/<session-id>.json

Default Workflow

Deep Distill (servers, recommended)

python ~/.codex/skills/manhua/session-distill/bin/deep-distill-run.py --offset 0 --batch-size 3

Then: answer-me verify → promote ANSWERED only → session-knowledge-base.md §Codex → check-work → mark distilled.

See references/deep-distill-workflow.md.

Legacy single-session flow

1. Run python ~/.codex/skills/manhua/session-distill/bin/session-distill.py run --next 1.
2. Read the generated packet and inspect Packet Audit first.
3. If Coverage: partial, inspect the raw JSONL span before promoting any conclusion.
4. Write a session note under distilled/sessions/<session-id>.md.
5. Append only stable reusable knowledge to knowledge-base.md, or explicitly record no promotion.
6. Before marking distilled, run a lightweight knowledge review when you promoted anything: python ~/.codex/skills/manhua/session-distill/bin/session-distill.py review-kb --next 20 --query <session-id-or-topic>
7. If memory drafts exist, review every entry so none remain pending.
8. Run python ~/.codex/skills/manhua/session-distill/bin/session-distill.py mark <session-id> distilled. This retains the original Codex rollout JSONL. If it is ever necessary to remove raw data, use the separate audited prune-raw <session-id> --confirm command with a reason.

The mark distilled command has guardrails: it fails unless the note exists, partial packets mention raw review, memory drafts have no pending entries, and the note records a promotion decision. If knowledge-base.md already contains entries sourced from the same session, obvious volatile or weak entries are also rejected unless you use --force. Raw-file deletion still only applies under ~/.codex/archived_sessions or ~/.codex/sessions.

Commands

python ~/.codex/skills/manhua/session-distill/bin/deep-distill-run.py --offset 0 --batch-size 3
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py status
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py list --size 100
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py run --next 1
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py mark <session-id> distilled
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py prune-raw <session-id> --confirm --reason "retention policy"
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py review-kb --next 20
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py verify-entry <session-id|keyword>
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py prune-kb --statuses stale,superseded
python ~/.codex/skills/manhua/session-distill/bin/session-distill.py self-test

References

• references/codex-session-format.md: Codex JSONL event shapes and packet rules.
• references/deep-distill-workflow.md: Grok-paradigm batch pipeline for servers.
• references/distillation-rules.md: promotion and filtering rules.
• references/output-layout.md: workspace layout and status meanings.

Use the older Claude session distiller only for ~/.claude/projects/*.jsonl.